Abstract
Slips, trips, and falls (STFs) are a major occupational hazard that contributes significantly to workplace injuries and the associated financial costs. The application of traditional fall detection techniques in the real world is limited because they are usually based on simulated falls. By using kinematic data from real near-fall incidents that occurred in physically demanding work environments, this study overcomes this limitation and improves the ecological validity of fall detection algorithms. This study systematically tests several machine-learning architectures for near-fall detection using the Prev-Fall dataset, which consists of high-resolution inertial measurement unit (IMU) data from 110 workers. Convolutional neural networks (CNNs), residual networks (ResNets), convolutional long short-term memory networks (convLSTMs), and InceptionTime models were trained and evaluated over a range of temporal window lengths using a neural architecture search. High-validation F1 scores were achieved by the best-performing models, particularly CNNs and InceptionTime, indicating their effectiveness in near-fall classification. The need for more contextual variables to increase robustness was highlighted by recurrent false positives found in subsequent tests on previously unobserved occupational data, especially during biomechanically demanding activities such as bending and squatting. Nevertheless, our findings suggest the applicability of machine-learning-based STF prevention systems for workplace safety monitoring and, more generally, applications in fall mitigation. To further improve the accuracy and generalizability of the system, future research should investigate multimodal data integration and improved classification techniques.